[Paper Review] Exploration of models for a cargo assembly planning problem

Gleb Belov, Natashia Boland+2 more|arXiv (Cornell University)|Apr 2, 2015

Optimization and Packing Problems18 references3 citations

TL;DR

This paper proposes a Constraint Programming (CP) approach with Large Neighbourhood Search (LNS) to minimize vessel delay in a coal stockyard cargo assembly problem, modeling dynamic scheduling under limited visibility horizons. The CP method outperforms an adaptive greedy heuristic in solution quality and scalability, especially with larger visibility horizons and extended constraints.

ABSTRACT

We consider a real-world cargo assembly planning problem arising in a coal supply chain. The cargoes are built on the stockyard at a port terminal from coal delivered by trains. Then the cargoes are loaded onto vessels. Only a limited number of arriving vessels is known in advance. The goal is to minimize the average delay time of the vessels over a long planning period. We model the problem in the MiniZinc constraint programming language and design a large neighbourhood search scheme. The effects of various optional constraints are investigated. Some of the optional constraints expand the model's scope toward a system view (berth capacity, port channel). An adaptive scheme for a greedy heuristic from the literature is proposed and compared to the constraint programming approach.

Motivation & Objective

• Address the real-world challenge of minimizing vessel delay in coal supply chain logistics at the Port of Newcastle.
• Develop a scalable and flexible planning model for cargo assembly in a stockyard with dynamic vessel arrivals and limited advance visibility.
• Compare a CP-based approach with a state-of-the-art greedy heuristic, improving solution quality and robustness.
• Investigate the impact of visibility horizon size, resource constraints (e.g., reclaimers), and operational parameters (e.g., stacking speed) on scheduling performance.
• Enable easy extension of the model to include new constraints, such as berth capacity and port channel limitations, for a holistic system view.

Proposed method

• Model the cargo assembly and vessel scheduling problem using MiniZinc with Constraint Programming (CP), capturing time-dependent stacking and reclaiming operations.
• Implement a Large Neighbourhood Search (LNS) scheme to iteratively improve solutions by exploring large, structured neighborhoods of the solution space.
• Introduce an adaptive iterative scheme for the greedy heuristic from prior work to enable fair comparison with the CP approach.
• Incorporate visibility horizon as a key parameter—controlling how many future vessels’ arrival and demand data are known in advance.
• Apply post-processing dwell reduction to minimize idle time between stockpile assembly and vessel loading.
• Extend the core model to include system-level constraints such as berth capacity, port channel limits, and reclaimer movement time.

Experimental results

Research questions

• RQ1 How does increasing the visibility horizon (number of known future vessels) affect the quality of vessel delay minimization solutions?
• RQ2 Can a CP-based approach with LNS outperform an existing greedy heuristic with partial lookahead in terms of solution quality and computational efficiency?
• RQ3 What is the impact of key operational constraints—such as limited reclaimers, stacking speed, and reclaimer movement time—on overall system performance?
• RQ4 How do extended model components (e.g., berth capacity, port channel) influence the feasibility and quality of solutions?
• RQ5 To what extent can the CP model be extended to include new constraints while maintaining scalability and solution quality?

Key findings

• A visibility horizon of 25 vessels produces solutions close to the best found, indicating diminishing returns beyond this point.
• Restricting reclaimers to three active at once increased average vessel delay by 91%, highlighting their critical role in system capacity.
• Increasing stacking speed by one day reduced delays by 65% in the basic model and by 11% when all constraints were included.
• The CP approach with LNS generated better overall solutions in less time than the adaptive greedy heuristic, especially with larger visibility horizons.
• Post-processing dwell reduction reduced average dwell time from 81 hours to 6.3 hours (100 vessels), significantly improving operational efficiency.
• Queue jumps (vessels processed out of ETA order) were observed, with CP solutions showing higher order permutations than the greedy heuristic, suggesting potential for future constraint-based mitigation.